AI’s most consequential developments this morning are about trust and the cost of scale, not just new models. Reporting published November 3 and early November 4 spans security warnings at OpenAI, AMD’s surging data-center business, Meta’s borrowing for infrastructure, a European copyright test for Gemini, and a prototype that could let customers verify which model answered them.
1. OpenAI faces renewed questions about its internal security
Reuters reported, citing five people familiar with the matter, that employees had warned leaders and directors for years about broad access to code and research, insufficient review of internal tools, and insider risks. The account follows the October disclosure of a recording device in an executive conference room; a former engineer was charged in connection with that incident. These are allegations and reported employee concerns, not a finding that model weights were exposed.
OpenAI disputes the idea that its most sensitive assets were generally accessible: it says model weights and other critical work are kept in separately controlled repositories and that it now reviews internal tools regularly. According to Reuters, executive Thibault Sottiaux reassigned a quarter of production engineers in August to security infrastructure. Spokesperson Elana Widmann said protecting the company’s technology, data and people is “a matter of national importance.” The episode illustrates the trade-off between an open research culture and tighter controls as frontier models become strategic assets.
2. AMD’s AI sales accelerate, but customer warrants carry a cost
AMD’s third-quarter data-center revenue climbed 97% to $8.55 billion, above the $8.20 billion analyst estimate, while company revenue reached $13.38 billion. Chief executive Lisa Su said on Tuesday that overall revenue could grow more than 80% in 2027; AMD forecast roughly $17 billion of fourth-quarter sales, plus or minus 3%. Its Helios AI rack, using the MI455X accelerator and Venice server chips, is in production, with early systems under test at Meta, Reuters reported.
The growth has an accounting and ownership dimension. AMD booked a $1.07 billion non-cash third-quarter charge, equivalent to 66 cents a share, tied to stock warrants offered to OpenAI and Meta as part of chip-supply agreements, Reuters separately reported. Meta’s agreement covers up to $60 billion of chips over five years. If commercial milestones are met, the warrants could ultimately dilute existing shareholders; finance chief Jean Hu said more guidance on their timing is expected in 2027. Strong AI demand is real, but the structure of the deals matters as much as the headline sales.
3. Meta turns to a $15 billion bond sale to fund AI expansion
Meta said Tuesday it was raising $15 billion in a five-part US dollar bond offering to finance its AI buildout. The tranche details include $3.25 billion due in 2031 at a 5.85% coupon and another $3.25 billion due in 2033 at 6.25%, according to Reuters’ account of the filing. The company has forecast $135 billion–$145 billion in capital expenditure this year, largely driven by AI, and had $83.67 billion in debt at September 30.
Moody’s assigned the new debt an Aa3 rating and said it expected Meta to retain a positive net cash position. The borrowing shows how even cash-rich platforms are leaning on capital markets to pay for data centers and compute. Investors will be watching whether eventual AI revenue justifies infrastructure commitments made today.
4. EU court adviser draws a line between AI training and copying news
In a case involving Hungarian publisher Like Company and Google’s Gemini, Court of Justice of the European Union Advocate General Maciej Szpunar said Tuesday that a chatbot may infringe publishers’ rights if it retrieves articles and reproduces their recognisable expression, or bypasses technical barriers. His non-binding opinion, reported by Reuters, distinguished protected expression from unprotected facts and said AI training on publicly available material can fall within the EU text-and-data-mining exception. He said there was no evidence in this case that Google trained Gemini on the publisher’s articles.
Google said its systems generate original outputs and pointed to publisher controls. The court will rule in the coming months; no infringement judgment has yet been issued. The distinction between training and answers that substitute for a publisher’s reporting could have significant consequences for AI search products and licensing negotiations.
5. Anthropic publishes FlashProve model-verification prototype
Anthropic has released FlashProve, open-source research into cryptographically proving that a model produced a particular response without revealing its weights. The goal is to let customers and auditors check that an advertised model—not an undisclosed substitute—served a request. TechCrunch reported Tuesday that the work includes code, technical documentation and formal-verification components, but is not slated for deployment.
Anthropic says a single H100 can generate proofs for a quantized 400-billion-parameter Llama 4 Maverick model at about 110 tokens per second, with a roughly 19 MB proof for a 1,024-token request. The company also says proof generation takes roughly 70–95 times the GPU time of ordinary inference in its comparison. Those substantial overheads, along with the current 1,024-token context limit, make this an intermediate research result rather than an off-the-shelf guarantee for everyday chat. If the approach matures, verifiable model identity could become useful for procurement, independent evaluation and auditing.
Together, the five stories show an industry moving beyond the question of what AI can do toward harder questions about who can verify it, secure it, finance it and govern its use.
